How attention can create synaptic tags for the learning of working memories in sequential tasks

A Circuit Model for Working Memory Based on Hybrid Positive and Negative-Derivative Feedback Mechanism

Working memory (WM) plays an important role in cognitive activity. The WM system is used to temporarily store information in learning and decision-making. WM always functions in many aspects of daily life, such as the short-term memory of words, cell phone verification codes, and cell phone numbers. In young adults, studies have shown that a central memory store is limited to three to five meaningful items. Little is known about how WM functions at the microscopic neural level, but appropriate neural network computational models can help us gain a better understanding of it. In this study, we attempt to design a microscopic neural network model to explain the internal mechanism of WM. The performance of existing positive feedback models depends on the parameters of a synapse. We use a negative-derivative feedback mechanism to counteract the drift in persistent activity, making the hybrid positive and negative-derivative feedback (HPNF) model more robust to common disturbances. To fulfill the mechanism of WM at the neural circuit level, we construct two main neural networks based on the HPNF model: a memory-storage sub-network (the memory-storage sub-network is composed of several sets of neurons, so we call it "SET network", or "SET" for short) with positive feedback and negative-derivative feedback and a storage distribution network (SDN) designed by combining SET for memory item storage and memory updating. The SET network is a neural information self-sustaining mechanism, which is robust to common disturbances; the SDN constructs a storage distribution network at the neural circuit level; the experimental results show that our network can fulfill the storage, association, updating, and forgetting of information at the level of neural circuits, and it can work in different individuals with little change in parameters.

Prediction-error neurons in circuits with multiple neuron types: Formation, refinement, and functional implications

An influential idea in neuroscience is that neural circuits do not only passively process sensory information but rather actively compare them with predictions thereof. A core element of this comparison is prediction-error neurons, the activity of which only changes upon mismatches between actual and predicted sensory stimuli. While it has been shown that these prediction-error neurons come in different variants, it is largely unresolved how they are simultaneously formed and shaped by highly interconnected neural networks. By using a computational model, we study the circuit-level mechanisms that give rise to different variants of prediction-error neurons. Our results shed light on the formation, refinement, and robustness of prediction-error circuits, an important step toward a better understanding of predictive processing.

Predictable sensory stimuli do not evoke significant responses in a subset of cortical excitatory neurons. Some of those neurons, however, change their activity upon mismatches between actual and predicted stimuli. Different variants of these prediction-error neurons exist, and they differ in their responses to unexpected sensory stimuli. However, it is unclear how these variants can develop and coexist in the same recurrent network and how they are simultaneously shaped by the astonishing diversity of inhibitory interneurons. Here, we study these questions in a computational network model with three types of inhibitory interneurons. We find that balancing excitation and inhibition in multiple pathways gives rise to heterogeneous prediction-error circuits. Dependent on the network’s initial connectivity and distribution of actual and predicted sensory inputs, these circuits can form different variants of prediction-error neurons that are robust to network perturbations and generalize to stimuli not seen during learning. These variants can be learned simultaneously via homeostatic inhibitory plasticity with low baseline firing rates. Finally, we demonstrate that prediction-error neurons can support biased perception, we illustrate a number of functional implications, and we discuss testable predictions.

How Working Memory and Reinforcement Learning Are Intertwined: A Cognitive, Neural, and Computational Perspective

Reinforcement learning and working memory are two core processes of human cognition and are often considered cognitively, neuroscientifically, and algorithmically distinct. Here, we show that the brain networks that support them actually overlap significantly and that they are less distinct cognitive processes than often assumed. We review literature demonstrating the benefits of considering each process to explain properties of the other and highlight recent work investigating their more complex interactions. We discuss how future research in both computational and cognitive sciences can benefit from one another, suggesting that a key missing piece for artificial agents to learn to behave with more human-like efficiency is taking working memory's role in learning seriously. This review highlights the risks of neglecting the interplay between different processes when studying human behavior (in particular when considering individual differences). We emphasize the importance of investigating these dynamics to build a comprehensive understanding of human cognition.

Learning at Variable Attentional Load Requires Cooperation of Working Memory, Meta-learning, and Attention-augmented Reinforcement Learning

Flexible learning of changing reward contingencies can be realized with different strategies. A fast learning strategy involves using working memory of recently rewarded objects to guide choices. A slower learning strategy uses prediction errors to gradually update value expectations to improve choices. How the fast and slow strategies work together in scenarios with real-world stimulus complexity is not well known. Here, we aim to disentangle their relative contributions in rhesus monkeys while they learned the relevance of object features at variable attentional load. We found that learning behavior across six monkeys is consistently best predicted with a model combining (i) fast working memory and (ii) slower reinforcement learning from differently weighted positive and negative prediction errors as well as (iii) selective suppression of nonchosen feature values and (iv) a meta-learning mechanism that enhances exploration rates based on a memory trace of recent errors. The optimal model parameter settings suggest that these mechanisms cooperate differently at low and high attentional loads. Whereas working memory was essential for efficient learning at lower attentional loads, enhanced weighting of negative prediction errors and meta-learning were essential for efficient learning at higher attentional loads. Together, these findings pinpoint a canonical set of learning mechanisms and suggest how they may cooperate when subjects flexibly adjust to environments with variable real-world attentional demands.

Circuit mechanisms for cortical plasticity and learning

The cerebral cortex integrates sensory information with emotional states and internal representations to produce coherent percepts, form associations, and execute voluntary actions. For the cortex to optimize perception, its neuronal network needs to dynamically retrieve and encode new information. Over the last few decades, research has started to provide insight into how the cortex serves these functions. Building on classical Hebbian plasticity models, the latest hypotheses hold that throughout experience and learning, streams of feedforward, feedback, and modulatory information operate in selective and coordinated manners to alter the strength of synapses and ultimately change the response properties of cortical neurons. Here, we describe cortical plasticity mechanisms that involve the concerted action of feedforward and long-range feedback input onto pyramidal neurons as well as the implication of local disinhibitory circuit motifs in this process.

A brain-inspired computational model for spatio-temporal information processing

Spatio-temporal information processing is fundamental in both brain functions and AI applications. Current strategies for spatio-temporal pattern recognition usually involve explicit feature extraction followed by feature aggregation, which requires a large amount of labeled data. In the present study, motivated by the subcortical visual pathway and early stages of the auditory pathway for motion and sound processing, we propose a novel brain-inspired computational model for generic spatio-temporal pattern recognition. The model consists of two modules, a reservoir module and a decision-making module. The former projects complex spatio-temporal patterns into spatially separated neural representations via its recurrent dynamics, the latter reads out neural representations via integrating information over time, and the two modules are linked together using known examples. Using synthetic data, we demonstrate that the model can extract the frequency and order information of temporal inputs. We apply the model to reproduce the looming pattern discrimination behavior as observed in experiments successfully. Furthermore, we apply the model to the gait recognition task, and demonstrate that our model accomplishes the recognition in an event-based manner and outperforms deep learning counterparts when training data is limited.

Efficient Spike-Driven Learning With Dendritic Event-Based Processing

A critical challenge in neuromorphic computing is to present computationally efficient algorithms of learning. When implementing gradient-based learning, error information must be routed through the network, such that each neuron knows its contribution to output, and thus how to adjust its weight. This is known as the credit assignment problem. Exactly implementing a solution like backpropagation involves weight sharing, which requires additional bandwidth and computations in a neuromorphic system. Instead, models of learning from neuroscience can provide inspiration for how to communicate error information efficiently, without weight sharing. Here we present a novel dendritic event-based processing (DEP) algorithm, using a two-compartment leaky integrate-and-fire neuron with partially segregated dendrites that effectively solves the credit assignment problem. In order to optimize the proposed algorithm, a dynamic fixed-point representation method and piecewise linear approximation approach are presented, while the synaptic events are binarized during learning. The presented optimization makes the proposed DEP algorithm very suitable for implementation in digital or mixed-signal neuromorphic hardware. The experimental results show that spiking representations can rapidly learn, achieving high performance by using the proposed DEP algorithm. We find the learning capability is affected by the degree of dendritic segregation, and the form of synaptic feedback connections. This study provides a bridge between the biological learning and neuromorphic learning, and is meaningful for the real-time applications in the field of artificial intelligence.

Modifying the m6A brain methylome by ALKBH5-mediated demethylation: a new contender for synaptic tagging

Synaptic plasticity processes, which underlie learning and memory formation, require RNA to be translated local to synapses. The synaptic tagging hypothesis has previously been proposed to explain how mRNAs are available at specific activated synapses. However how RNA is regulated, and which transcripts are silenced or processed as part of the tagging process is still unknown. Modification of RNA by N6-methyladenosine (m⁶A/m) influences the cellular fate of mRNA. Here, by advanced microscopy, we showed that m⁶A demethylation by the eraser protein ALKBH5 occurs at active synaptic ribosomes and at synapses during short term plasticity. We demonstrated that at activated glutamatergic post-synaptic sites, both the YTHDF1 and YTHDF3 reader and the ALKBH5 eraser proteins increase in co-localisation to m⁶A-modified RNAs; but only the readers showed high co-localisation to modified RNAs during late-stage plasticity. The YTHDF1 and YTHFDF3 readers also exhibited differential roles during synaptic maturation suggesting that temporal and subcellular abundance may determine specific function. m⁶A-sequencing of human parahippocampus brain tissue revealed distinct white and grey matter m⁶A methylome profiles indicating that cellular context is a fundamental factor dictating regulated pathways. However, in both neuronal and glial cell-rich tissue, m⁶A effector proteins are themselves modified and m⁶A epitranscriptional and posttranslational modification processes coregulate protein cascades. We hypothesise that the availability m⁶A effector protein machinery in conjunction with RNA modification, may be important in the formation of condensed synaptic nanodomain assemblies through liquid-liquid phase separation. Our findings support that m⁶A demethylation by ALKBH5 is an intrinsic component of the synaptic tagging hypothesis and a molecular switch which leads to alterations in the RNA methylome, synaptic dysfunction and potentially reversible disease states.

Flexible Working Memory Through Selective Gating and Attentional Tagging

Working memory is essential: it serves to guide intelligent behavior of humans and nonhuman primates when task-relevant stimuli are no longer present to the senses. Moreover, complex tasks often require that multiple working memory representations can be flexibly and independently maintained, prioritized, and updated according to changing task demands. Thus far, neural network models of working memory have been unable to offer an integrative account of how such control mechanisms can be acquired in a biologically plausible manner. Here, we present WorkMATe, a neural network architecture that models cognitive control over working memory content and learns the appropriate control operations needed to solve complex working memory tasks. Key components of the model include a gated memory circuit that is controlled by internal actions, encoding sensory information through untrained connections, and a neural circuit that matches sensory inputs to memory content. The network is trained by means of a biologically plausible reinforcement learning rule that relies on attentional feedback and reward prediction errors to guide synaptic updates. We demonstrate that the model successfully acquires policies to solve classical working memory tasks, such as delayed recognition and delayed pro-saccade/anti-saccade tasks. In addition, the model solves much more complex tasks, including the hierarchical 12-AX task or the ABAB ordered recognition task, both of which demand an agent to independently store and updated multiple items separately in memory. Furthermore, the control strategies that the model acquires for these tasks subsequently generalize to new task contexts with novel stimuli, thus bringing symbolic production rule qualities to a neural network architecture. As such, WorkMATe provides a new solution for the neural implementation of flexible memory control.

A functional spiking-neuron model of activity-silent working memory in humans based on calcium-mediated short-term synaptic plasticity

In this paper, we present a functional spiking-neuron model of human working memory (WM). This model combines neural firing for encoding of information with activity-silent maintenance. While it used to be widely assumed that information in WM is maintained through persistent recurrent activity, recent studies have shown that information can be maintained without persistent firing; instead, information can be stored in activity-silent states. A candidate mechanism underlying this type of storage is short-term synaptic plasticity (STSP), by which the strength of connections between neurons rapidly changes to encode new information. To demonstrate that STSP can lead to functional behavior, we integrated STSP by means of calcium-mediated synaptic facilitation in a large-scale spiking-neuron model and added a decision mechanism. The model was used to simulate a recent study that measured behavior and EEG activity of participants in three delayed-response tasks. In these tasks, one or two visual gratings had to be maintained in WM, and compared to subsequent probes. The original study demonstrated that WM contents and its priority status could be decoded from neural activity elicited by a task-irrelevant stimulus displayed during the activity-silent maintenance period. In support of our model, we show that it can perform these tasks, and that both its behavior as well as its neural representations are in agreement with the human data. We conclude that information in WM can be effectively maintained in activity-silent states by means of calcium-mediated STSP.

Mentally maintaining information for short periods of time in working memory is crucial for human adaptive behavior. It was recently shown that the human brain does not only store information through neural firing–as was widely believed–but also maintains information in activity-silent states. Here, we present a detailed neural model of how this could happen in our brain through short-term synaptic plasticity: rapidly adapting the connection strengths between neurons in response to incoming information. By reactivating the adapted network, the stored information can be read out later. We show that our model can perform three working memory tasks as accurately as human participants can, while using similar mental representations. We conclude that our model is a plausible and effective neural implementation of human working memory.

Dissociation of nicotinic α7 and α4/β2 sub-receptor agonists for enhancing learning and attentional filtering in nonhuman primates

RATIONALE

Nicotinic acetylcholine receptors (nAChRs) modulate attention, memory, and higher executive functioning, but it is unclear how nACh sub-receptors mediate different mechanisms supporting these functions.

OBJECTIVES

We investigated whether selective agonists for the alpha-7 nAChR versus the alpha-4/beta-2 nAChR have unique functional contributions for value learning and attentional filtering of distractors in the nonhuman primate.

METHODS

Two adult rhesus macaque monkeys performed reversal learning following systemic administration of either the alpha-7 nAChR agonist PHA-543613 or the alpha-4/beta-2 nAChR agonist ABT-089 or a vehicle control. Behavioral analysis quantified performance accuracy, speed of processing, reversal learning speed, the control of distractor interference, perseveration tendencies, and motivation.

RESULTS

We found that the alpha-7 nAChR agonist PHA-543613 enhanced the learning speed of feature values but did not modulate how salient distracting information was filtered from ongoing choice processes. In contrast, the selective alpha-4/beta-2 nAChR agonist ABT-089 did not affect learning speed but reduced distractibility. This dissociation was dose-dependent and evident in the absence of systematic changes in overall performance, reward intake, motivation to perform the task, perseveration tendencies, or reaction times.

CONCLUSIONS

These results suggest nicotinic sub-receptor specific mechanisms consistent with (1) alpha-4/beta-2 nAChR specific amplification of cholinergic transients in prefrontal cortex linked to enhanced cue detection in light of interferences, and (2) alpha-7 nAChR specific activation prolonging cholinergic transients, which could facilitate subjects to follow-through with newly established attentional strategies when outcome contingencies change. These insights will be critical for developing function-specific drugs alleviating attention and learning deficits in neuro-psychiatric diseases.

Conscious Processing and the Global Neuronal Workspace Hypothesis

We review the central tenets and neuroanatomical basis of the global neuronal workspace (GNW) hypothesis, which attempts to account for the main scientific observations regarding the elementary mechanisms of conscious processing in the human brain. The GNW hypothesis proposes that, in the conscious state, a non-linear network ignition associated with recurrent processing amplifies and sustains a neural representation, allowing the corresponding information to be globally accessed by local processors. We examine this hypothesis in light of recent data that contrast brain activity evoked by either conscious or non-conscious contents, as well as during conscious or non-conscious states, particularly general anesthesia. We also discuss the relationship between the intertwined concepts of conscious processing, attention, and working memory.

Control of synaptic plasticity in deep cortical networks

Humans and many other animals have an enormous capacity to learn about sensory stimuli and to master new skills. However, many of the mechanisms that enable us to learn remain to be understood. One of the greatest challenges of systems neuroscience is to explain how synaptic connections change to support maximally adaptive behaviour. Here, we provide an overview of factors that determine the change in the strength of synapses, with a focus on synaptic plasticity in sensory cortices. We review the influence of neuromodulators and feedback connections in synaptic plasticity and suggest a specific framework in which these factors can interact to improve the functioning of the entire network.

A Closed-Loop Toolchain for Neural Network Simulations of Learning Autonomous Agents

Neural network simulation is an important tool for generating and evaluating hypotheses on the structure, dynamics, and function of neural circuits. For scientific questions addressing organisms operating autonomously in their environments, in particular where learning is involved, it is crucial to be able to operate such simulations in a closed-loop fashion. In such a set-up, the neural agent continuously receives sensory stimuli from the environment and provides motor signals that manipulate the environment or move the agent within it. So far, most studies requiring such functionality have been conducted with custom simulation scripts and manually implemented tasks. This makes it difficult for other researchers to reproduce and build upon previous work and nearly impossible to compare the performance of different learning architectures. In this work, we present a novel approach to solve this problem, connecting benchmark tools from the field of machine learning and state-of-the-art neural network simulators from computational neuroscience. The resulting toolchain enables researchers in both fields to make use of well-tested high-performance simulation software supporting biologically plausible neuron, synapse and network models and allows them to evaluate and compare their approach on the basis of standardized environments with various levels of complexity. We demonstrate the functionality of the toolchain by implementing a neuronal actor-critic architecture for reinforcement learning in the NEST simulator and successfully training it on two different environments from the OpenAI Gym. We compare its performance to a previously suggested neural network model of reinforcement learning in the basal ganglia and a generic Q-learning algorithm.

An Integrative Framework for Sensory, Motor, and Cognitive Functions of the Posterior Parietal Cortex

Throughout the history of modern neuroscience, the parietal cortex has been associated with a wide array of sensory, motor, and cognitive functions. The use of non-human primates as a model organism has been instrumental in our current understanding of how areas in the posterior parietal cortex (PPC) modulate our perception and influence our behavior. In this Perspective, we highlight a series of influential studies over the last five decades examining the role of the PPC in visual perception and motor planning. We also integrate long-standing views of PPC functions with more recent evidence to propose a more general model framework to explain integrative sensory, motor, and cognitive functions of the PPC.

Biologically plausible deep learning - But how far can we go with shallow networks?

Training deep neural networks with the error backpropagation algorithm is considered implausible from a biological perspective. Numerous recent publications suggest elaborate models for biologically plausible variants of deep learning, typically defining success as reaching around 98% test accuracy on the MNIST data set. Here, we investigate how far we can go on digit (MNIST) and object (CIFAR10) classification with biologically plausible, local learning rules in a network with one hidden layer and a single readout layer. The hidden layer weights are either fixed (random or random Gabor filters) or trained with unsupervised methods (Principal/Independent Component Analysis or Sparse Coding) that can be implemented by local learning rules. The readout layer is trained with a supervised, local learning rule. We first implement these models with rate neurons. This comparison reveals, first, that unsupervised learning does not lead to better performance than fixed random projections or Gabor filters for large hidden layers. Second, networks with localized receptive fields perform significantly better than networks with all-to-all connectivity and can reach backpropagation performance on MNIST. We then implement two of the networks - fixed, localized, random & random Gabor filters in the hidden layer - with spiking leaky integrate-and-fire neurons and spike timing dependent plasticity to train the readout layer. These spiking models achieve >98.2% test accuracy on MNIST, which is close to the performance of rate networks with one hidden layer trained with backpropagation. The performance of our shallow network models is comparable to most current biologically plausible models of deep learning. Furthermore, our results with a shallow spiking network provide an important reference and suggest the use of data sets other than MNIST for testing the performance of future models of biologically plausible deep learning.

Feature-specific prediction errors and surprise across macaque fronto-striatal circuits

To adjust expectations efficiently, prediction errors need to be associated with the precise features that gave rise to the unexpected outcome, but this credit assignment may be problematic if stimuli differ on multiple dimensions and it is ambiguous which feature dimension caused the outcome. Here, we report a potential solution: neurons in four recorded areas of the anterior fronto-striatal networks encode prediction errors that are specific to feature values of different dimensions of attended multidimensional stimuli. The most ubiquitous prediction error occurred for the reward-relevant dimension. Feature-specific prediction error signals a) emerge on average shortly after non-specific prediction error signals, b) arise earliest in the anterior cingulate cortex and later in dorsolateral prefrontal cortex, caudate and ventral striatum, and c) contribute to feature-based stimulus selection after learning. Thus, a widely-distributed feature-specific eligibility trace may be used to update synaptic weights for improved feature-based attention.

In order to adjust expectations efficiently, prediction errors need to be associated with the features that gave rise to the unexpected outcome. Here, the authors show that neurons in anterior fronto-striatal networks encode prediction errors that are specific to feature values of different stimulus dimensions.

Dendritic solutions to the credit assignment problem

Guaranteeing that synaptic plasticity leads to effective learning requires a means for assigning credit to each neuron for its contribution to behavior. The 'credit assignment problem' refers to the fact that credit assignment is non-trivial in hierarchical networks with multiple stages of processing. One difficulty is that if credit signals are integrated with other inputs, then it is hard for synaptic plasticity rules to distinguish credit-related activity from non-credit-related activity. A potential solution is to use the spatial layout and non-linear properties of dendrites to distinguish credit signals from other inputs. In cortical pyramidal neurons, evidence hints that top-down feedback signals are integrated in the distal apical dendrites and have a distinct impact on spike-firing and synaptic plasticity. This suggests that the distal apical dendrites of pyramidal neurons help the brain to solve the credit assignment problem.

Multi-Timescale Memory Dynamics Extend Task Repertoire in a Reinforcement Learning Network With Attention-Gated Memory

The interplay of reinforcement learning and memory is at the core of several recent neural network models, such as the Attention-Gated MEmory Tagging (AuGMEnT) model. While successful at various animal learning tasks, we find that the AuGMEnT network is unable to cope with some hierarchical tasks, where higher-level stimuli have to be maintained over a long time, while lower-level stimuli need to be remembered and forgotten over a shorter timescale. To overcome this limitation, we introduce a hybrid AuGMEnT, with leaky (or short-timescale) and non-leaky (or long-timescale) memory units, that allows the exchange of low-level information while maintaining high-level one. We test the performance of the hybrid AuGMEnT network on two cognitive reference tasks, sequence prediction and 12AX.

Eligibility Traces and Plasticity on Behavioral Time Scales: Experimental Support of NeoHebbian Three-Factor Learning Rules

Most elementary behaviors such as moving the arm to grasp an object or walking into the next room to explore a museum evolve on the time scale of seconds; in contrast, neuronal action potentials occur on the time scale of a few milliseconds. Learning rules of the brain must therefore bridge the gap between these two different time scales. Modern theories of synaptic plasticity have postulated that the co-activation of pre- and postsynaptic neurons sets a flag at the synapse, called an eligibility trace, that leads to a weight change only if an additional factor is present while the flag is set. This third factor, signaling reward, punishment, surprise, or novelty, could be implemented by the phasic activity of neuromodulators or specific neuronal inputs signaling special events. While the theoretical framework has been developed over the last decades, experimental evidence in support of eligibility traces on the time scale of seconds has been collected only during the last few years. Here we review, in the context of three-factor rules of synaptic plasticity, four key experiments that support the role of synaptic eligibility traces in combination with a third factor as a biological implementation of neoHebbian three-factor learning rules.

Contextual signals in visual cortex

Vision is an active process. What we perceive strongly depends on our actions, intentions and expectations. During visual processing, these internal signals therefore need to be integrated with the visual information from the retina. The mechanisms of how this is achieved by the visual system are still poorly understood. Advances in recording and manipulating neuronal activity in specific cell types and axonal projections together with tools for circuit tracing are beginning to shed light on the neuronal circuit mechanisms of how internal, contextual signals shape sensory representations. Here we review recent work, primarily in mice, that has advanced our understanding of these processes, focusing on contextual signals related to locomotion, behavioural relevance and predictions.

The influence of attention and reward on the learning of stimulus-response associations

We can learn new tasks by listening to a teacher, but we can also learn by trial-and-error. Here, we investigate the factors that determine how participants learn new stimulus-response mappings by trial-and-error. Does learning in human observers comply with reinforcement learning theories, which describe how subjects learn from rewards and punishments? If yes, what is the influence of selective attention in the learning process? We developed a novel redundant-relevant learning paradigm to examine the conjoint influence of attention and reward feedback. We found that subjects only learned stimulus-response mappings for attended shapes, even when unattended shapes were equally informative. Reward magnitude also influenced learning, an effect that was stronger for attended than for non-attended shapes and that carried over to a subsequent visual search task. Our results provide insights into how attention and reward jointly determine how we learn. They support the powerful learning rules that capitalize on the conjoint influence of these two factors on neuronal plasticity.

The Distributed Nature of Working Memory

Studies in humans and non-human primates have provided evidence for storage of working memory contents in multiple regions ranging from sensory to parietal and prefrontal cortex. We discuss potential explanations for these distributed representations: (i) features in sensory regions versus prefrontal cortex differ in the level of abstractness and generalizability; and (ii) features in prefrontal cortex reflect representations that are transformed for guidance of upcoming behavioral actions. We propose that the propensity to produce persistent activity is a general feature of cortical networks. Future studies may have to shift focus from asking where working memory can be observed in the brain to how a range of specialized brain areas together transform sensory information into a delayed behavioral response.

The laminar integration of sensory inputs with feedback signals in human cortex

•

Understanding how the cortex integrates feedback and feedforward signals is central to understanding brain function.

•

The data-driven framework of apical amplification which is hypothesized to have a central role in cognition is highlighted.

•

Human neuroimaging data provides evidence for layer-specific cortical feedback relevant for theories of predictive feedback.

The cortex constitutes the largest area of the human brain. Yet we have only a basic understanding of how the cortex performs one vital function: the integration of sensory signals (carried by feedforward pathways) with internal representations (carried by feedback pathways). A multi-scale, multi-species approach is essential for understanding the site of integration, computational mechanism and functional role of this processing. To improve our knowledge we must rely on brain imaging with improved spatial and temporal resolution and paradigms which can measure internal processes in the human brain, and on the bridging of disciplines in order to characterize this processing at cellular and circuit levels. We highlight apical amplification as one potential mechanism for integrating feedforward and feedback inputs within pyramidal neurons in the rodent brain. We reflect on the challenges and progress in applying this model neuronal process to the study of human cognition. We conclude that cortical-layer specific measures in humans will be an essential contribution for better understanding the landscape of information in cortical feedback, helping to bridge the explanatory gap.

Toward an Integration of Deep Learning and Neuroscience

Neuroscience has focused on the detailed implementation of computation, studying neural codes, dynamics and circuits. In machine learning, however, artificial neural networks tend to eschew precisely designed codes, dynamics or circuits in favor of brute force optimization of a cost function, often using simple and relatively uniform initial architectures. Two recent developments have emerged within machine learning that create an opportunity to connect these seemingly divergent perspectives. First, structured architectures are used, including dedicated systems for attention, recursion and various forms of short- and long-term memory storage. Second, cost functions and training procedures have become more complex and are varied across layers and over time. Here we think about the brain in terms of these ideas. We hypothesize that (1) the brain optimizes cost functions, (2) the cost functions are diverse and differ across brain locations and over development, and (3) optimization operates within a pre-structured architecture matched to the computational problems posed by behavior. In support of these hypotheses, we argue that a range of implementations of credit assignment through multiple layers of neurons are compatible with our current knowledge of neural circuitry, and that the brain's specialized systems can be interpreted as enabling efficient optimization for specific problem classes. Such a heterogeneously optimized system, enabled by a series of interacting cost functions, serves to make learning data-efficient and precisely targeted to the needs of the organism. We suggest directions by which neuroscience could seek to refine and test these hypotheses.

Learning a New Selection Rule in Visual and Frontal Cortex

How do you make a decision if you do not know the rules of the game? Models of sensory decision-making suggest that choices are slow if evidence is weak, but they may only apply if the subject knows the task rules. Here, we asked how the learning of a new rule influences neuronal activity in the visual (area V1) and frontal cortex (area FEF) of monkeys. We devised a new icon-selection task. On each day, the monkeys saw 2 new icons (small pictures) and learned which one was relevant. We rewarded eye movements to a saccade target connected to the relevant icon with a curve. Neurons in visual and frontal cortex coded the monkey's choice, because the representation of the selected curve was enhanced. Learning delayed the neuronal selection signals and we uncovered the cause of this delay in V1, where learning to select the relevant icon caused an early suppression of surrounding image elements. These results demonstrate that the learning of a new rule causes a transition from fast and random decisions to a more considerate strategy that takes additional time and they reveal the contribution of visual and frontal cortex to the learning process.

Reinforcement Learning of Linking and Tracing Contours in Recurrent Neural Networks

The processing of a visual stimulus can be subdivided into a number of stages. Upon stimulus presentation there is an early phase of feedforward processing where the visual information is propagated from lower to higher visual areas for the extraction of basic and complex stimulus features. This is followed by a later phase where horizontal connections within areas and feedback connections from higher areas back to lower areas come into play. In this later phase, image elements that are behaviorally relevant are grouped by Gestalt grouping rules and are labeled in the cortex with enhanced neuronal activity (object-based attention in psychology). Recent neurophysiological studies revealed that reward-based learning influences these recurrent grouping processes, but it is not well understood how rewards train recurrent circuits for perceptual organization. This paper examines the mechanisms for reward-based learning of new grouping rules. We derive a learning rule that can explain how rewards influence the information flow through feedforward, horizontal and feedback connections. We illustrate the efficiency with two tasks that have been used to study the neuronal correlates of perceptual organization in early visual cortex. The first task is called contour-integration and demands the integration of collinear contour elements into an elongated curve. We show how reward-based learning causes an enhancement of the representation of the to-be-grouped elements at early levels of a recurrent neural network, just as is observed in the visual cortex of monkeys. The second task is curve-tracing where the aim is to determine the endpoint of an elongated curve composed of connected image elements. If trained with the new learning rule, neural networks learn to propagate enhanced activity over the curve, in accordance with neurophysiological data. We close the paper with a number of model predictions that can be tested in future neurophysiological and computational studies.

Our experience with the visual world allows us to group image elements that belong to the same perceptual object and to segregate them from other objects and the background. If subjects learn to group contour elements, this experience influences neuronal activity in early visual cortical areas, including the primary visual cortex (V1). Learning presumably depends on alterations in the pattern of connections within and between areas of the visual cortex. However, the processes that control changes in connectivity are not well understood. Here we present the first computational model that can train a neural network to integrate collinear contour elements into elongated curves and to trace a curve through the visual field. The new learning algorithm trains fully recurrent neural networks, provided the connectivity causes the networks to reach a stable state. The model reproduces the behavioral performance of monkeys trained in these tasks and explains the patterns of neuronal activity in the visual cortex that emerge during learning, which is remarkable because the only feedback for the model is a reward for successful trials. We discuss a number of the model predictions that can be tested in future neuroscientific work.

Implicit Value Updating Explains Transitive Inference Performance: The Betasort Model

Transitive inference (the ability to infer that B > D given that B > C and C > D) is a widespread characteristic of serial learning, observed in dozens of species. Despite these robust behavioral effects, reinforcement learning models reliant on reward prediction error or associative strength routinely fail to perform these inferences. We propose an algorithm called betasort, inspired by cognitive processes, which performs transitive inference at low computational cost. This is accomplished by (1) representing stimulus positions along a unit span using beta distributions, (2) treating positive and negative feedback asymmetrically, and (3) updating the position of every stimulus during every trial, whether that stimulus was visible or not. Performance was compared for rhesus macaques, humans, and the betasort algorithm, as well as Q-learning, an established reward-prediction error (RPE) model. Of these, only Q-learning failed to respond above chance during critical test trials. Betasort’s success (when compared to RPE models) and its computational efficiency (when compared to full Markov decision process implementations) suggests that the study of reinforcement learning in organisms will be best served by a feature-driven approach to comparing formal models.

Although machine learning systems can solve a wide variety of problems, they remain limited in their ability to make logical inferences. We developed a new computational model, called betasort, which addresses these limitations for a certain class of problems: Those in which the algorithm must infer the order of a set of items by trial and error. Unlike extant machine learning systems (but like children and many non-human animals), betasort is able to perform “transitive inferences” about the ordering of a set of images. The patterns of error made by betasort resemble those made by children and non-human animals, and the resulting learning achieved at low computational cost. Additionally, betasort is difficult to classify as either “model-free” or “model-based” according to the formal specifications of those classifications in the machine learning literature. One of the broader implications of these results is that achieving a more comprehensive understanding of how the brain learns will require analysts to entertain other candidate learning models.

Spatial diversity of spontaneous activity in the cortex

The neocortex is a layered sheet across which a basic organization is thought to widely apply. The variety of spontaneous activity patterns is similar throughout the cortex, consistent with the notion of a basic cortical organization. However, the basic organization is only an outline which needs adjustments and additions to account for the structural and functional diversity across cortical layers and areas. Such diversity suggests that spontaneous activity is spatially diverse in any particular behavioral state. Accordingly, this review summarizes the laminar and areal diversity in cortical activity during fixation and slow oscillations, and the effects of attention, anesthesia and plasticity on the cortical distribution of spontaneous activity. Among questions that remain open, characterizing the spatial diversity in spontaneous membrane potential may help elucidate how differences in circuitry among cortical regions supports their varied functions. More work is also needed to understand whether cortical spontaneous activity not only reflects cortical circuitry, but also contributes to determining the outcome of plasticity, so that it is itself a factor shaping the functional diversity of the cortex.